“They only care to show us the wheelchair”: disability representation in text-to-image AI models

Title of the Paper

“They only care to show us the wheelchair”: Disability Representation in Text-to-Image AI Models

Bibliographic Information

• Research Area: Human-Computer Interaction, Generative Artificial Intelligence, and Social Justice
• Keywords: Disability Representation, Generative AI, Algorithmic Bias, AI Ethics, Text-to-Image Models, Social Justice

Research Background and Problem Statement

• What issues or challenges did the authors identify? The paper highlights significant biases in current text-to-image (T2I) generative AI models when representing disabled communities. These biases include oversimplified, stereotypical, or negative portrayals, such as excessive focus on wheelchairs, depictions of sadness and loneliness, and a lack of diversity in age, gender, and race.

• Why is this issue important? Disabled communities have long been marginalized in mainstream media and societal representations, reinforcing harmful stereotypes and potentially affecting their mental health and social inclusion. The widespread adoption of generative AI increases the risk of perpetuating and amplifying these misrepresentations, posing challenges for the disabled community.

• Research Motivation and Related Work The authors build on prior findings that AI systems exhibit systemic biases, including those inherited by emerging technologies like text-to-image generation. However, research on disability representation remains limited and predominantly focused on language models. This study aims to address this gap by systematically investigating the impact of T2I models on disabled communities and setting higher representation standards for generative AI.

Proposed Solutions

• What methods or solutions did the authors propose?

1. Investigating how current text-to-image models represent disabled individuals, including identifying forms and causes of bias.
2. Analyzing systemic errors in image generation technology and providing improvement recommendations, such as more accurate training datasets, diversity-conscious output mechanisms, and enhanced user interfaces.

• What is innovative about these solutions? The study employs focus groups and qualitative analysis as innovative methods, involving 25 disabled participants to provide community-centered feedback on AI representation practices. The authors also propose multidimensional improvements, spanning model design to image generation processes, while addressing potential ethical conflicts.

• What are the implementation steps and key technologies used?

1. Conducting focus groups to collect user reactions to AI-generated images, with eight focus group studies completed.
2. Generating images for various disability types and cultural contexts to analyze potential errors and biases.
3. Using qualitative content analysis to code focus group data and extract specific representation suggestions from participants.

Research Results

• What specific results were achieved?

1. Identified recurring stereotypes in text-to-image models’ disability representations, such as excessive portrayal of wheelchairs and inaccurate or outdated depictions of assistive technologies.
2. Summarized disabled individuals’ demands for improved representation, including more diverse outputs and more nuanced depictions of emotions and settings.

• How does this compare to existing solutions? Unlike previous studies relying on automated evaluations, this research incorporates in-depth feedback through community participation, highlighting unresolved issues in generative AI’s disability representation and offering detailed technical and design recommendations.

• What were the experimental or evaluation results? The study revealed that current models systematically generate inaccurate and negative representations, failing to address the diverse needs of disabled communities. Participants emphasized the need for AI to better understand and reflect real-world disability experiences.

• Limitations and Future Directions

1. The study focuses only on disabled communities in North America, potentially lacking a global perspective.
2. Text-to-image generative AI technology is still evolving, and inherent biases in data may not be immediately resolved through model improvements. Future research directions include:

1. Developing cross-cultural, multilingual datasets for disability representation;
2. Exploring value-driven AI design to balance authenticity, respect, and diversity in representations;
3. Conducting global community engagement studies to include disabled populations from diverse regions and cultural backgrounds.

Recommended Improvements

Image Characteristics

• Increase diversity in race, gender, and age depicted in images.
• Avoid inaccurate or outdated portrayals of assistive technologies, while focusing on their everyday applications.

User Interface Suggestions

• Provide detailed explanations of generated images to help users understand the intended representation purposes.
• Incorporate metadata tagging related to disability to better represent invisible disabilities and avoid overemphasizing restrictive features of disabled individuals.

System Development

• Involve disabled individuals in model evaluation and development processes.
• Implement community-trust-based model design practices, including collecting rich and diverse training datasets, particularly for assistive technologies and disability cultural symbols.

Conclusion

This paper identifies multiple shortcomings in generative AI models’ disability representations and offers a series of innovative suggestions to promote fairness from both technical and ethical perspectives. The findings provide valuable insights for establishing systemic fairness metrics and design processes in future AI development.